The Agent Skills Directory

[COMMAND_EXECUTION] (SAFE): Uses standard filesystem commands (find, mkdir, mv, test) to reorganize documentation. These operations are restricted to the local environment and the specific task of file management.
[DATA_EXFILTRATION] (SAFE): The skill reads project metadata (README, package.json) and documentation files to categorize content. No network-based exfiltration patterns or external URLs were detected.
[PROMPT_INJECTION] (LOW): This skill is subject to indirect prompt injection (Category 8).
Ingestion points: Reads first 100 lines of all .md files in the /docs directory, as well as root configuration files.
Boundary markers: None. The agent is instructed to read and categorize content without specific delimiters or instructions to ignore embedded commands.
Capability inventory: Can move files, create directories, and modify specific project files like CLAUDE.md.
Sanitization: None. The agent processes raw text from existing documentation.
Assessment: A malicious actor could place instructions inside a markdown file (e.g., in a hidden comment) to influence how the agent reorganizes the documentation or what it writes into the blueprint.md file. However, the impact is limited to the documentation scope.

docs-builder